跳到正文
SakanaAI/ShinkaEvolve · 提交动态· Robert Tjarko Lange·· 2 天前AI 评分30

SakanaAI ShinkaEvolve 优化 worker 池并发上限并记录每次 clamp(#235)

perf(runner): limit each worker pool by what its workers cost, and log every clamp (#235) Concurrency follow-up to the async speedup check. - Proposal and database workers are no longer cut by core count: proposals mostly wait on an LLM, database workers are threads. Evaluation jobs stay limited by cores and memory. - Slurm evaluation jobs are no longer limited by the machine that submits them. - Every reduction is logged at WARNING with the requested and granted numbers. - The async guide and README state measured scaling instead of a 5-10x claim. Measured with a stub LLM on a 10-core machine: 1, 2 and 4 workers unchanged; 8/8/4 workers about +15 to 20 percent at 3 s LLM latency and +70 percent at 15 s. Independently reviewed; the review's findings are fixed in this branch. Raises default CPU use and concurrent LLM requests for configurations the old rule cut.

来源与原文

SakanaAI/ShinkaEvolve · 提交动态

作者:Robert Tjarko Lange

发布时间:(北京时间)

阅读原始来源 · github.com ↗
完整链接与内容说明

https://github.com/SakanaAI/ShinkaEvolve/commit/2c6b6e38cde6c20ec0acff1d129f792efdae5e36

本站目前仅有 AI 导读与摘要,未展示原文正文;技术细节、代码和完整结论请查看原始来源。

AI 导读

SakanaAI 的 ShinkaEvolve 项目合并了 PR #235:proposal 和 database workers 不再按 CPU 核数削减,Slurm 评测任务不再受提交机器限制,每次上限削减都会以 WARNING 级日志记录请求数和实际授予数。实测(stub LLM、10 核机器)显示 8 个 worker 在 3 秒 LLM 延迟下提速约 15-20%,15 秒延迟下提速约 70%;async 指南和 README 改用实测数据替代原来的 5-10x 说法。

来源:SakanaAI/ShinkaEvolve · 提交动态 · github.com